Re: [patch] x86: allow ZONE_DMA to be configurable

From: Andrew Morton
Date: Thu Oct 14 2010 - 18:33:26 EST


On Thu, 14 Oct 2010 15:15:28 -0700 (PDT)
David Rientjes <rientjes@xxxxxxxxxx> wrote:

> On Thu, 14 Oct 2010, Andrew Morton wrote:
>
> > > ZONE_DMA is unnecessary for a large number of machines that do not
> > > require addressing in the lower 16MB of memory because they do not use
> > > ISA devices with 16-bit address registers (plus one page byte register).
> > >
> > > This patch allows users to disable ZONE_DMA for x86 if they know they
> > > will not be using such devices with their kernel.
> > >
> > > This prevents the VM from unnecessarily reserving a ratio of memory
> > > (defaulting to 1/256th of system capacity) with lowmem_reserve_ratio
> > > for such allocations when it will never be used.
> > >
> >
> > I wonder how hard it would be to do this at runtime, probably with a
> > boot parameter.
> >
>
> A "no_zone_dma" boot parameter wouldn't allow us to achieve the text or
> data savings that we see from disabling all three options in question
> here.

True, but doing it at runtime is vastly more user-friendly. Distros
aren't doing to double the number of kernels they package for this
option!

Of course, we could do both.

> Hot-adding ZONE_DMA at runtime would be possible but there's no guarantee
> that memory hasn't fully been used by the time you do it, so it disregards
> lowmem_reserve_ratio unless you migrate everything, which has a dependency
> on it being movable.
>
> > I'd be a little concerned at the effects of this on page reclaim and
> > the page allocator - it might expose weird pre-existing bugs or
> > inefficiencies. But we can cross that bridge when we fall off it, I
> > guess.
> >
>
> We've run with it for a couple years, we can even undefine __GFP_DMA to
> find allocations that we compile into the kernel to ensure we don't have a
> requirement for the zone.

"we" is google, I assume.

> Perhaps only define the gfp flag when we have
> CONFIG_ZONE_DMA and break users' builds until they disable options that
> require it

That sounds a good idea.

What happens in the current patch if someone passes __GFP_DMA when
CONFIG_ZONE_DMA=n? ENOMEM? Perhaps it should WARN() as well.

> (or enable CONFIG_ZONE_DMA)?

hm. It'd be better to make all the drivers depend on ZONE_DMA. Good
luck with that :)

> It would be great if we could do "select ZONE_DMA" for all options that
> require it, though.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/